IC-318CLIP backbones from different architectures (ViTs and ResNets) trained with the same data and objective exhibit complementary strengths, with an oracle per-image backbone selection improving zero-shot accuracy by up to 43.5% over the best single backbone
Cristian Rodriguez-Opazo, Ehsan Abbasnejad, Damien Teney, Hamed Damirchi, Edison Marrese-Taylor, Anton van den Hengel
The paper evaluates 9 released CLIP backbones (5 ResNets, 4 ViTs) on 21 zero-shot classification datasets. Despite identical pre-training data and objective, the backbones correctly classify different subsets of images. On ImageNet-1k, ViT-L/14-336 gets 38,285 correct while RN50x64 gets 36,963, yet the oracle combining all backbones reaches 44,506. The oracle improvement is largest on fine-grained or domain-specific datasets: 43.5% on EuroSAT, 36.0% on CLEVR, 25.1% on CUB, 21.8% on DTD. The paper also notes that within the same family, larger models do not always outperform smaller ones (e.g., ViT-B/32 beats RN50x4 on CIFAR-10 and CIFAR-100), and that the diversity of predictions across backbones strongly correlates with the achievable ensemble gain.
Evidence
correlational
Key metric
oracle all improvements over best single backbone: 43.5% (EuroSAT), 36.0% (CLEVR), 25.1% (CUB), 21.8% (DTD), 16.6% (CIFAR-100), 16.0% (Cars); ImageNet-1k: ViT-L/14-336 38,285, RN50x64 36,963, oracle all 44,506 out of 50,000
Caveat
The oracle is an upper bound that assumes perfect per-image backbone selection; the practical NLC method achieves 9.1% average improvement, not 43.5%. Evaluation is limited to zero-shot classification on 21 datasets.